How Self-Published Rankings Reach AI Answers — and What China's 315 Exposé Means for European Companies

In our own cross-platform testing this September, we kept seeing a pattern in AI answers about providers in this space: a ranked list, a precise-looking score, "2026 rankings" language that reads like independent evaluation. In at least one case the ranking traced back to a page published by the company it ranked first. China's state broadcaster CCTV had exposed the underlying mechanism on national television in March 2026, and a peer-reviewed position paper has since named how it works.

We have written before about why China's regulatory layer shapes what AI platforms treat as a citable source[3]. This piece asks a related question: what happens when content is built to be cited, and can the systems retrieving it tell the difference?

Why are the stakes so high?

FACT

China's market for generative engine optimization (GEO) services, the whole industry of optimizing content for AI answers, legitimate and questionable alike, was estimated by iiMedia Research at CNY 34.9 billion (about USD 5.1 billion) in 2025, up 67% on the year[2][7]. That is one research firm's estimate, and it measures the entire industry, not manipulation. What it shows is how much commercial value now depends on being the source an AI system cites.

INFERENCE

Where that much value depends on one outcome, some participants will try to manufacture it.

Is the manipulation itself documented, or only suspected?

FACT

Documented, on national television. CCTV's annual consumer-rights broadcast (315 Gala, 15 March 2026) described services that mass-produce content across the internet to interfere with how AI models collect and generate answers, sold openly with tiered pricing on major domestic e-commerce platforms[1][6].

What, mechanically, makes this work?

FACT

A position paper presented at this year's International Conference on Machine Learning names two techniques: "retrieval booster messages," designed to get content pulled into an AI-generated answer at all, and "ranking shifter messages," designed to shape how it is framed once included[4]. The same paper cites Microsoft's own documentation of hidden prompts designed to steer AI assistants toward particular companies, and an OECD-recorded case of models recommending fictitious or low-quality products.

INFERENCE

Our reading: a numbered list, named categories and a decimal score look like independent evaluation to a retrieval system, largely regardless of who wrote them.

Who is accountable — the platform, or whoever publishes the content?

INFERENCE, grounded in on-record commentary

Xinhua's coverage of the broadcast quoted Liao Huaixue, a partner at Tahota Law Firm, stating that "the brand is the initiator and ultimate beneficiary" of such services[5]. That places real responsibility on the publisher, not only on the platform that surfaces it. Yao Jinxin, a researcher at the Wuzhen Institute and a senior member of the Institute of Electrical and Electronics Engineers (IEEE), described the mechanism not as rewriting the AI's brain but as "contaminating the materials it consults," and warned that a platform's credibility suffers when commercial content passes as a neutral answer[5]. That points to the platform side: today's retrieval systems mostly lack a reliable way to tell an independent evaluation from a party evaluating itself. Both are true, and neither excuses the other.

Does it show up in ordinary use?

FACT

In our September baseline audit, we put a set of neutral, unprompted questions to five AI platforms (ChatGPT, Claude.ai, Gemini, Perplexity and Grok), repeating each on three separate days. On questions about providers for European companies entering China, the answers regularly included ranked lists with numeric scores. Handling differed by platform. Perplexity repeated vendor-published scores as ranking data, without qualification, in both its first and third runs. Claude.ai described the sector's rankings as self-promotional in several of its runs, and ChatGPT did so in some of its web-enabled runs.

INFERENCE

Nothing about this kind of content reliably stops an AI system from citing it as fact, and nothing reliably catches it either: whether a reader gets a caveat looks more like a matter of which platform they ask, and sometimes which run they get, than of the content itself. One limit: as of this writing, our testing covered Western platforms. How Chinese platforms treat the same pattern is something we have not yet measured.

Where is the line between legitimate and manufactured?

INFERENCE

Not everything in GEO is manipulation. The useful distinction is whether a claim can be checked from outside the party making it. A ranking published on an independent host, with a stated methodology and a disclosed stake, is a different object from a templated "Top 10" page on the ranked company's own domain, even when both look identical inside an AI answer. The first can be verified. The second only asks to be believed.

What does this mean for a European company?

Three risks follow from the evidence above. They are not equally well supported, so each is labeled.

Risk in choosing a provider

FACT

In our baseline audit, AI answers about China-side providers included ranked lists with scores that some platforms repeated without a caveat. A shortlist built that way can rest on rankings the listed companies wrote themselves.

INFERENCE

How a provider markets itself is the first sample of how it will treat your content. The signal is disclosure and independence, not self-publication as such: a ranking labeled as the company's own view is a different thing from a templated page presented as independent.

Risk from the provider's tactics, and who answers for them

FACT

The 315 broadcast exposed a specific GEO system, and at least two providers publicly distanced themselves from manipulation in the days that followed[8][9].

INFERENCE

The legal expert quoted by Xinhua called the brand the "initiator and ultimate beneficiary" of such services[5]. It is reasonable to expect that brands that commission such content, not only the providers, could come under scrutiny. This is a reading of the evidence, not a prediction.

Risk to the result and its durability

FACT

AI answers vary between runs[10], so a provider's claim to have "got you cited" based on a single query is not evidence. China's market regulator has also listed AI-generated advertising as a 2026 supervision priority[8].

INFERENCE

A tactic that depends on manufactured content may stop working, or become a liability, if platforms or regulators act on it; platforms accountable for their answers have reason to prefer sources they can defend[3]. We have not tested how Chinese platforms or regulators respond, so this too is a reading, not a prediction.

What should a European company do with an AI's "top ranked" answer?

Treat a ranking inside an AI answer like a restaurant that tops a list published on its own website: first ask who wrote the list. Three checks travel well. Does the ranking live only on the ranked party's own site, with no independent host or verification anywhere in the chain? Is the same company's name attached to an unusually large number of separately labeled "top" or "best" categories, which tends to indicate a template built to multiply pages rather than a single considered evaluation? And does the publisher disclose its own stake plainly, or does the conflict only become visible once you check who owns the domain?

When hiring a provider, add one more question: how does it produce and source its content, and is any ranking it cites independent?

Related insights

All insights & research →

About the author

Marin Tintchev is the founder of GeoSynapse and China GEO EU. He has 30+ years in ICT and international business across China and Europe, was educated at Tsinghua University, and is fluent in Mandarin.

Connect on LinkedIn →